• Article
  • October 3, 2018

Using Predictive Analytics to Reduce Hospital Readmissions

This hospital uses data to assess the likelihood of readmission.

In the broadest sense, pediatric medicine has used data for years to improve patient outcomes. Standardized reports, data visualization dashboards, patient registries and other forms of reporting have—and will continue to be—important tools for clinicians. But what William Feaster and his team at Children's Hospital of Orange County (CHOC) are doing now takes data to a new level.

"We wanted to go beyond just reporting data to providing true knowledge about a patient based on the data," says Feaster, M.D., CHOC's chief health information officer. "Using advanced data science tools, we're able to use data on thousands of patients to predict an outcome for one given patient based on all the patients we'd seen before."

Predictive model targets patient readmissions

Despite efforts to reduce readmissions within seven days of patient discharge, CHOC's readmission rates were consistently in line with averages experienced by their peers. Feaster believed there was room for improvement, so he—along with Louis Ehwerhemuepha, Ph.D., a data scientist at CHOC, and their team—developed a predictive model aimed at helping clinicians better anticipate patients at risk for readmission.

To do that, the team looked at four years of data—roughly 38,000 patient discharges. Applying sophisticated algorithms and hundreds of variables to this data, Feaster and his team built a tool that ultimately has achieved an accuracy of 0.82 area under the curve (AUC), which Feaster describes as "highly predictive."

Using this model, Feaster's team assesses the likelihood of readmission for every patient at CHOC (not including those in intensive care, who are evaluated on a separate scale), and assigns a risk factor of high, medium or low. These assessments are shared with the care managers, who can use the information to guide their discharge procedures. Although the risk assessments don't prescribe specific patient interventions, the six primary risk areas they provide for each patient helps the clinical teams focus on the most likely areas of concern.

"We are leaving it up to the clinical staff to determine what's most important for that patient," Feaster says. "Patient care is still done at the bedside and is customized for every patient. But if you can alert people to risk that a particular patient has, you can figure out what issues may lead to that patient being readmitted."

Building models in a "super-charged environment"

CHOC is only a few months in to the implementation phase of this predictive model, so Feaster says it's too early to determine if the patient assessments are having an effect on readmission rates. One thing that is certain—the rapid advancement of this technology will speed more predictive tools into practice.

"It took us about six months to build a complete model on this, but that's changing very quickly," Feaster says. "You can now get a solution to your problem in a matter of minutes—something that might have taken weeks to run previously. We're now working in this super-charged environment, so developing algorithms to answer clinical questions will be much quicker and much easier." As a result, CHOC currently has several projects in the works to apply data to various aspects of medical care, according to Feaster.

Data and the future

Feaster says feedback from the case managers has been positive. "They're very excited about it," he says. CHOC plans to integrate the information gleaned from the predictive model into patient electronic medical records soon, so a patient's entire care team will have access to the risk assessment and be able to adjust care accordingly.

What excites Feaster most is what this burgeoning technology means for the future of pediatric medicine. "How we practice medicine is going to be dramatically enhanced," Feaster says. "This is going to take off and absolutely explode over the next five to 10 years."

Send questions or comments.